Explicit Baseball Event Detection by Combining Visual and Speech Information

To explicitly detect baseball events, a multimodal approach that combines the decisions of visual and speech event detection is proposed. We respectively perform event detection from visual and speech perspectives and estimate the corresponding confidences. By combining the decisions from different modalities, the detection performance increases by 8% ~ 20%, in terms of F1 metric. This work achieves explicit event detection in baseball videos and facilitates the development of realistic applications for management or entertainment.